Abstract

Apple fruitlet thinning is an important management practice each spring to ensure the production of a crop that balances yield, fruit size, and return bloom. The chemical thinner rate and timing is therefore one of the most important decisions a grower faces to get the desired abscision response and cropload. Having said that it is also a widely variable process that depends heavily on tree carbohydrate availability, which is difficult to quantify. There is currently only one resource available to growers to inform thinner rates (carbon balance model), it however is only based on simulations and not site specific measurements. This study aims to assess the plausability of using NDVI as a proxy for carbon balance throughout the thinning window and test any relationships between weather conditions and NDVI. NDVI was captured for eight days from the middle of April through the end of June for an orchard in Quincy, Washington. Initial results showed that NDVI did correspond well to anticipated carbohydrate status: NDVI was lowest shortly after bloom when there is intense competition for carbohydrates and increased as the season continued and fruit droped. Temperature and solar radiation either the day before or the day of NDVI measurements were not a good predictor of NDVI. K-means analysis showed consistent clustering in early season measurements and from mid-season on, the variance is therefore homogenous between measurement dates. K-means cluster of all dates together once again showed the trend of the later dates having higher NDVI measurements. The trends in this study show the promise in utilizing NDVI as an indicator of carbohydrate status and more indepth studies should be pursued.

Background and Driving Questions

Apple fruitlet thinning is an important annual cultural practice that growers use to manage a crop load that balances yield, fruit size, and bloom for the following year: high current season crop load can suppress flower buds that would bloom the next year. It is estimated that up to 70 – 90% of fruitlets need to abscise to reach this desired crop load, this is mainly done through chemical thinning applications. Abscission response to chemical thinning spray is quite variable and is dependent on tree carbohydrate status. Fruitlet abscission occurs during a small window of 4 – 6 weeks after bloom (thinning window), during a time when fruitlets are competing for carbohydrates between other fruitlets and developing shoots. The quantification of carbohydrates available to fruitlets is a function of the carbohyrates supplied to the tree through photosynthesis minus the carbohydrates demanded by sinks and consumed for respiration.

The more carbohydrates that are available for fruit growth, the less likely fruit are going to abscise and the higher the chemical thinner rate would need to be applied to get the desired thinning outcome. Conversely, trees that are under higher levels of carbohydrate stress are going to shed fruit more easily and should have a decreased rate of chemical thinner applied. Weather conditions are one of the main factors that can affect carbohydrate status. Conditions that promote photosynthesis and limit respiration are associated with higher carbohydrate availability: cool, but sunny days. Whereas cloudy and high temperatures result in low photosynthesis, but high respiration and lower carbohydrate availability.

Getting the desired percent of fruitlet abscission is a critical decision that growers are faced with and one that there are few resources available to determine correct rates. Remotely sensed multispectral imaging tree health indicators could be used as an estimate of tree carbohydrate status to better inform chemical thinning spray rates. Normalized difference vegetation index (NDVI) is one such remotely sensed measurement that quantifies the difference between near infrared reflectance and red light reflectance normalized to the sum of these two wavelengths; giving a value between ± 1. The closer the NDVI value is to one, the healthier the vegetation. This exploratory data analysis this study first aims to look at how NDVI levels in an orchard vary throughout the abscission window, specifically do expected carbohydrate stress levels correspond to NDVI? The second aim is to test any relationships these NDVI levels have with weather conditions.

Dataset Description

Satelitte data was downloaded from EarthExplorer for an orhchard site in Quincy, Washington (fig. 1) over the course of the spring and into the summer where a range of carbohydrate availability is expected. The original idea was to choose an orchard at the NC State Mountain Horticultural Crops Research and Extension Center; however, from April 1 - June 30, there were only three cloud free days where satellite data could be downloaded for. Therefore, an orchard in Wasington where there is tpically less cloud cover was chosen. Even at this site there was only a total of eight day for which NDVI data could be downloaded: April 15, 20, 30; May 8, 10, 28, 30; and June 24. NDVI data is downloaded as a .tif file and is converted to a data frame of raster data with an x and y location and the NDVI value.

Figure 1: Image of orchard near Quincy, WA

Figure 1: Image of orchard near Quincy, WA

Weather data was downloaded from AgWeatherNet through Washington State University at a weather station in the near vacinity of the orhcard. Weather was downloaded on an hourly basis from April 1 - June 30. Of this data that was downloaded, the variables that were kept were: date, time, min temperature, average temperature, maximum temperature, and solar radiation in w/m2. The average of the top 10 hour irradiance and temperature measurements were then calculated for the day of NDVI measurement and the day before. This data was then combined with the NDVI data frame to give the full dataset for analysis; these variables are summarized in table 1.

Table 1: dataset variables
Variable Name in dataset Type
Date date Date
Date in relation to NDVI capture day chr
Raster X location x num
Raster Y location y num
Normalized difference vegetation idex NDVI num
Solar Radiation irradiance num
Temperature temp num

Methods

Tidy Data

Raster data had to first be converted to a dataframe with an x, y location and the NDVI value. Since there were separate dataframes for each date, the variable date had to be added to the data and each date was merged into a single dataframe. Only positive NDVI values correspond to vegetation, so data was filtered to only include NDVI values greater than zero.

Tidying the weather data took a bit more work. First, the column names needed to be renamed so that they didn’t have units in them, symbols, or spaces. To only look at weather conditions affecting photosynthesis, any observations where irradiance was less than zero was removed. Since NDVI values were only for one snapshot of time on each of the days captured, it was tricky to match up weather conditons to those values. I decided to calculate the average of the 10 highest irradiance and temperature observations for the day or and the day before each NDVI measurement.The weather dataframe then had columns: date; irradiance day of, and before measurement; temperature day of and before measurement. This data was then joined with the NDVI data frame and gathered to have an observation for each x, y coordinate with temperature and irradiance on the day of or before NDVI was measured.

Visualizations

To first assess the NDVI values for the orchard, raster data was viewed for the full NDVI dataset; this was faceted by day to see how orchard values change over time (fig. 2). A boxplot was then then made to look at the variability of NDVI values for each day and see the overall trend of NDVI values throughout the thinning window (fig. 3). A scatterplot of actual NDVI observations was overlayed with the boxplot to better get a sense of the density of observations.

Spatial variability of the orchard was analysed through k-means clustering and these clusters were plotted in a scatterplot. I first plotted each day in a different facet to see how clusters changed with time (fig. 4). Each of these days were combined in figure 5, to view clusters for the whole dataset, differentiated by day. In each of these cluster plots, the scaled x, y coordinate was plotted on the x-axis and scaled NDVI values were plotted on the y-axis.

Statistical summaries

I chose to report the median NDVI values along with the average temperature and irradiance for the 10 highest hours on the day that NDVI was measured (table 2). I chose to report the median NDVI because from the boxplot there does seem to be some outliers that would influence the mean. For the weather conditions, I chose to take the average of the 10 highest observations because this should give the average conditions over the entire daylight time when the trees would be actively photosynthesizing. This table is meant to provide a snapshot of what is going on each day that NDVI was measured.

Statistical analysis

K-means clustering was done to analyze the spatial variability of the orchard and to determine if there were areas that had consistently higher or lower NDVI values. Orchard point locations were first determined by combining the x and y coordinates. To calculate clusters, both orchard location and NDVI were scaled: with the median value being zero at values above that being positive and below being negative. For each individual day, optimum number of clusters was determined by the “silhoutte” method. K-means were then calculated with the determined number of clusters and a specified 25 sets. K-means were calculated for each day and then for all days combined.

Multiple regression was done to see how irradiance, temperature and their interaction could be used to predict NDVI. A separate model was built for weather conditions the day before and day of NDVI measurement to see if there was a lag in NDVI by weather. To limit orchard variability of NDVI, this analysis was done on just one section of the orchard over time.

Results and Discussion

Upon first glance of raster NDVI plots it appears that expected carbohydte stress levels do correspond with NDVI values (fig. 2). Bloom in this growing region for 2019 close to the long term average of late april to early may. This means that carbodydrate stress would be most likely to occur in around early May, shortly after bloom. Over the whole orchard, NDVI values do look to be the lowest on April 30th, May 8th and 10th; falling in line with the timeline above. June 24th, a date outside of the thinning window, has higher NDVI values than any of the other dates. These visual trends are supported by boxplots of NDVI on each date (fig. 3), where there is a dip in NDVI on April 30th, May 8th and 10th; then a steady increase in NDVI until June 24th.

Figure 2: NDVI raster data by date for orchard near Quincy, WA

Figure 2: NDVI raster data by date for orchard near Quincy, WA

Figure 3: boxplot of NDVI values by date

Figure 3: boxplot of NDVI values by date

The table below summarizes the main data of the study for each date measured: median NDVI, average daytime temperature and irradiance. This summary does show the overall trend described above with the lowest NDVI values occuring in late April/early May - shortly after bloom. There are some interesting comparisons to be made with weather conditions with days that are close to eachother, mainly that conditions that promote photosynthesis are occuring with the higher NDVI date. For example comparing May 10th to May 8th and April 20th to April 15th there are higher NDVI values with higher irradiance in the latter days.

Table 2: summary of median NDVI values by day with average temperature and irradiance that day
Date NDVI Temperature Irradiance
April 15 0.1875000 53.11 558.5
April 20 0.2155010 66.27 625.3
April 30 0.1916550 61.57 655.1
May 8 0.1813380 74.82 501.1
May 10 0.2173913 78.10 651.5
May 28 0.2508335 80.18 626.6
May 30 0.2459016 82.13 660.7
June 24 0.2710706 72.80 728.8

K-means clusters shows that there is consistent changes in NDVI by location: NDVI by locations are clustering simalarly by dates (fig. 4). There does seem to be a bit of a change in clustering between the measurements in April and the rest of the dates. One possible explanation for this is that there are differences in bloom between areas of the orchard. This is a large orchard and there are most certainly different cultivars that would have varying bloom dates. In this hypothetical, variation that would be clustering together when there are differences in phenology would not be seen once the whole orchard has bloomed (May onward). The consistency of clusters after May 10th shows that variances are fairly homogenous variance between locations in the orchard.

Figure 4: k-means cluster for NDVI values by orchard location for each date

Figure 4: k-means cluster for NDVI values by orchard location for each date

Clusters of all the dates combined is further evidence that NDVI values are higher as carbohydrate competition lessens (fig. 5). While there are a lot of observations, there is a pretty clear spectrum of NDVI values by date, with the later dates: May 28th and 30th, and June 24th clustering higher than the earlier dates. Here you see many of the lowest observations being light orange color, belonging to either April 30th or May 8th.

Figure 5: k-means cluster for NDVI values by orchard location for all dates

Figure 5: k-means cluster for NDVI values by orchard location for all dates

Regression analysis showed that weather conditions either the day before or the day of were not very good predictors of NDVI; with r2 values of 0.0534 and 0.181, respectively (table 3 and 4). Weather conditions the day of NDVI measurements were a much better predictor of NDVI than conditions the day before. This points to the notion that if NDVI is capturing carbohydrate stress there isn’t much of a lag in the measurement from the weather conditions that are influencing photosynthesis and inturn carbohydrate status. The obvious fault in this analysis is that NDVI from satellites can only be measured on sunny days when there is little cloud cover. This means that we are unable to compare NDVI values on cloudy days, when trees have decreased photosynthesis and are more stressed, we may be able to see a better relationship between NDVI and conditions with a wider range of weather conditions.

Table 3: regression model for weather conditions day before NDVI measurement
term estimate std.error statistic p.value
(Intercept) 0.0905293 0.0186640 4.850466 0.0000013
irradiance 0.0000505 0.0000393 1.285409 0.1986809
temp 0.0017874 0.0003010 5.939314 0.0000000
irradiance:temp -0.0000007 0.0000006 -1.177496 0.2390275
Note: r-squared = 0.0534
Table 4: regression model for weather conditions day of NDVI measurement
term estimate std.error statistic p.value
(Intercept) 0.8320966 0.0566316 14.69314 0
irradiance -0.0011862 0.0000941 -12.61129 0
temp -0.0110111 0.0007862 -14.00554 0
irradiance:temp 0.0000203 0.0000013 15.62123 0
Note: r-squared = 0.181

This exploratory data analysis shows the promise of using NDVI, or another mutispectral imaging approach, to be a proxy for carbon balance quntification. There was an clear trend between NDVI and anticipated carbohydrate stress with the lowest NDVI values occuring shortly after bloom when trees are most stressed for carbohydrates, but increasing with decreasing carbohydrate demand as the growing season continues. Clearly, there are many factors that is not known about this orchard such as cultivars used, orchard age, and cultural management practices that prohibit any hard conclusions to be drawn. Furthermore, there were only 8 days out of a three month period that were able to be used for analyisis and we are unable to distinguish between trees and the ground cover in the row alleys with raster data; these are the limitations of satellite captured data. However, the results of this study show that more detailed work that captures NDVI by either unmanned aerial vehicles or ground based systems with a uniform interval and over a range of weather conditions is worthwhile. Ultimately, a model could be developed to aid growers in chemical thinning applications that is based on NDVI that accounts for orchard variability and is rooted in data actually measured in that orchard.